feature(sunjx): add GSPO and GMPO algorithms support

[Jiaxuan-Sun]

Implement GSPO and GMPO Algorithms for Policy Optimization

This PR implements GSPO and GMPO algorithms.

GSPO Reference Implementation: verl

Usage: Set --advantage_estimator "gspo" or --advantage_estimator "gmpo" in training scripts.

⚠️ Note: GSPO implementation follows the reference code, but initial training experiments show model performance does not yet align with baseline results. Further tuning may be required.

[puyuan1996]

请参考 https://github.com/verl-project/verl/blob/4612d6d16258c71b2e81295cee4f1a9f2ad803c4/verl/trainer/ppo/core_algos.py#L1254 和 https://github.com/callsys/GMPO 的实现